the future-fit tab

What if you bought a new machine?

Try the boxes people are actually talking about. See what their memory could hold before you spend real money.

01 · Pick the fantasy

Which machine are you imagining?

Model context

02 · Read the envelope

DGX Spark

A self-contained CUDA workstation for serious local inference.

128 GiBmodel budget ≈ 112 GiB

GB10 Grace Blackwell4 TB NVMe10 GbE + ConnectX-7

32 comfortable fitsat 16k · serious sessionsBuy now

Qwen3-Coder-Next-FP8CUDA · 8-bit · runtime lane verified

86.3 GiBComfortable

Qwen3-Coder-30B-A3B-Instruct-GPTQ-Int8CUDA · 8-bit · runtime lane verified

35.9 GiBComfortable

Qwen3-Coder-30B-A3B-Instruct-FP8CUDA · 8-bit · runtime lane verified

35.0 GiBComfortable

Qwen3.6-35B-A3B-AWQCUDA · 4-bit · runtime lane verified

28.4 GiBComfortable

Qwen2.5-VL-32B-Instruct-AWQCUDA · 4-bit · runtime lane verified

26.6 GiBComfortable

Qwen3.6-35B-A3B-GPTQ-Int4CUDA · 4-bit · runtime lane verified

25.5 GiBComfortable

Qwen2.5-32B-Instruct-GPTQ-Int4CUDA · 4-bit · runtime lane verified

25.2 GiBComfortable

Qwen2.5-32B-Instruct-AWQCUDA · 4-bit · runtime lane verified

25.2 GiBComfortable

Qwen2.5-Coder-32B-Instruct-AWQCUDA · 4-bit · runtime lane verified

25.2 GiBComfortable

Qwen3-32B-AWQCUDA · 4-bit · runtime lane verified

25.2 GiBComfortable

Qwen3.8-27B-GPTQ-INT4-FP8KVCUDA · COMPRESSED-TENSORS · runtime lane verified

23.5 GiBComfortable

Qwopus3.6-27B-Coder-NVFP4CUDA · COMPRESSED-TENSORS · runtime lane verified

22.1 GiBComfortable

View DGX Spark Check manufacturer specs

Checked against the saved LocalModelMatch catalog on Sep 15, 2026. These are memory estimates, not performance tests or purchase advice. Prices and availability can change.